Small immunological clocks identified by deep learning and gradient boosting.

Area of Science:

• Immunology
• Computational Biology
• Artificial Intelligence

Background:

• Aging is linked to increased inflammation and immune system changes.
• These age-related immune alterations can drive disease and systemic inflammation.

Purpose of the Study:

• To develop a simplified model for predicting "immunological age" using a minimal set of biomarkers.
• To leverage advanced machine learning techniques for accurate age prediction from immunological data.

Main Methods:

• Utilized Elastic Net, gradient-boosted decision trees, and deep neural networks (DANet, SAINT, FT-Transformer, TabNet) to model chronological age from cytokine data.
• Applied SHAP values for dimensionality reduction to identify key age-associated immunological parameters.
• Constructed the SImAge clock using the top 10 identified immunological parameters.

Main Results:

• The FT-Transformer deep neural network achieved the best performance for the SImAge model.
• Achieved a mean absolute error of 6.94 years and a Pearson correlation coefficient (ρ) of 0.939 on an independent test dataset.
• Explainable AI methods provided individual participant-level model explanations.

Conclusions:

• Successfully developed the SImAge model, an immunological age clock based on 10 parameters, with FT-Transformer as the optimal deep learning architecture.
• SImAge demonstrates competitive accuracy with fewer input features compared to existing immunological profile studies.
• Deep neural networks show superior performance over gradient-boosted trees for immunological profile analysis and are recommended for future research.